Meta releases Llama 4 and the open vs closed source debate gets real

What Meta released

The Llama 4 family included Scout, Maverick, and a preview of a model called Behemoth. Scout is designed for long-context tasks with a 10-million-token context window. Maverick is the general-purpose model, competitive with GPT-4o on many benchmarks. Behemoth is described as Meta's frontier model, a 288-billion active parameter mixture-of-experts system that was still in training at release time.

All models were released as downloadable weights, meaning you can run them on your own hardware without sending data to Meta's servers. The models are also free for commercial use for most organisations. The license does restrict use by companies with over 700 million monthly active users, which effectively means the only entity that cannot freely use Llama 4 commercially is a competitor at the scale of Google or Microsoft.

Is Llama actually open source?

This argument comes up every Llama release and it is worth engaging with directly. The Open Source Initiative defines open source software as including the right to study, modify, distribute, and use the source for any purpose. Llama's license restricts commercial use for large organisations and does not release training code, datasets, or the full methodology used to produce the models.

By the OSI definition, Llama is not open source. It is open weights, which is meaningfully different. You can run the model and modify the weights through fine-tuning, but you cannot reproduce the training run, audit the data, or build a competing product at the same scale without significant independent investment.

Meta uses "open source AI" in its marketing. This frustrates some researchers and developers who care about the distinction. But practically, open weights delivers most of the benefits that developers actually want: local deployment, privacy, fine-tuning, no API costs, no rate limits.

Open weights without training data is like receiving a car without the factory blueprints. You can drive it, repaint it, even tune the engine. You cannot build your own car from first principles using what you received.

Why Meta keeps giving away frontier models

The strategic logic for Meta is reasonably clear. Meta's core business is social media advertising. It does not sell AI services directly. Every dollar a developer spends on OpenAI or Google API calls is a dollar that could be displaced by a free Meta model. Free models also generate adoption, goodwill, research collaborations, and fine-tuning work that flows back into the ecosystem Meta benefits from.

Additionally, Meta employs some of the best AI researchers in the world and spends billions on compute. Open-releasing models means the global research community spends its time improving and extending your base, finding your bugs, building applications on your infrastructure. The training cost is real, but the return in ecosystem development is hard to quantify and probably significant.

Open vs closed: the actual performance comparison

In early 2024, the best open-weights models lagged the best closed models by a significant margin. Llama 2 was useful but clearly behind GPT-4. By the time Llama 4 Maverick arrived in April 2025, that gap had compressed substantially on most standard benchmarks. Maverick performed comparably to GPT-4o and Claude 3.5 Sonnet on instruction-following, coding, and reasoning tasks.

The gap that remains is at the very top. OpenAI's o3, Anthropic's Claude 3.7 extended thinking, and Google's Gemini Ultra still outperform the open-weights models on the hardest reasoning benchmarks. But for most applications, the practical difference between the tiers has narrowed enough that the choice between open and closed is now genuinely a product decision rather than a capability one.

What this means for developers

The practical effect of Llama 4 is that local deployment of a genuinely capable frontier-class model became viable for any organisation with reasonable GPU infrastructure. Healthcare companies that cannot send patient data to external APIs can run a model that competes with GPT-4o. Financial institutions with strict data residency requirements have options. Researchers in countries with limited access to US API services can work with the same generation of model.

The competitive pressure this creates is real. OpenAI and Anthropic have to justify their API pricing relative to a free alternative. The answer they typically offer is reliability, safety investment, and enterprise support. That is a credible answer for some customers. For others, the economics of self-hosting are compelling enough that Llama 4 changes their calculus.